Got Personality

“Got Personality” – A cartoon that illustrates how AI models’ personalities are becoming visible in the wild.

We didn’t set out to give AI a personality, but we’re starting to put it in situations where personality matters. The moment models stopped answering exam questions and started making decisions, especially when money was on the line, their differences started looking like temperament.

A recent live trading experiment made this clear. Frontier models from OpenAI, Google, Anthropic, xAI, and Chinese labs like DeepSeek and Alibaba’s Qwen were given identical capital ($350K), data, and constraints. Over roughly two weeks, outcomes ranged from gains above 30% to losses exceeding 25%. What separated them wasn’t raw intelligence. It was how they reacted under pressure.

DeepSeek and Qwen traded with structure and conviction, closer to systematic or macro funds than reactive traders. They concentrated risk, adjusted positions mid-trade, and stuck with a thesis longer than most humans would. Google’s Gemini behaved in a way anyone who has watched retail trading would recognize … flipping positions after losses and hedging quickly after gains, driven by recent outcomes rather than structure. Anthropic’s Claude sat between the two. Thoughtful and patient, but occasionally undone by position sizing. OpenAI’s GPT models showed the sharpest evolution. Early versions overtraded and overleveraged. A later incremental release tightened risk controls and jumped from near the bottom of the table to near the top without any change in market conditions.

Even xAI’s Grok split into two personalities. One version leaned into volatility and imploded. Its successor imposed boundaries and finished as the strongest performer.

These behavior personalities are shaped by data, alignment, and incentives. As AI models continue to retrain and specialize, their personalities will diverge. Choosing an AI will feel less like picking software, and more like deciding who you want in the room when decisions get hard.